[PATCH/RFC] ARM64: use this_cpu_read in raw_smp_processor_id()

[Will Deacon]

On Thu, Dec 01, 2016 at 04:00:07PM +0100, Marek Szyprowski wrote:
> Direct access to cpu_number entry in per-cpu variables causes boot
> failure on Exynos5433, so replace it with this_cpu_read() macro.
> This approach is also used on x86_64.

Right, but x86 doesn't need to disable preemption in their per-cpu ops
afaik, so they don't take the performance hit.

Is this failure specific to Exynos5433?

> Signed-off-by: Marek Szyprowski <m.szyprowski at samsung.com>
> ---
> This change is needed to get linux-next to boot on Exynos5433, otherwise it
> hangs somewhere in early init. There is even no message on the earlycon.
> 
> This issue appeared first on linux-next from 14.11.2016. The tree from
> 11.11.2016 is the last one, which boots on Exynos5433. I've tried to
> debug a bit this issue, but I ran out of ideas.

I suspect the culprit is 57c82954e77f ("arm64: make cpu number a percpu
variable").
> 
> Any comments or suggestions are welcome.
> 
> Best regards
> Marek Szyprowski
> Samsung R&D Institute Poland
> ---
>  arch/arm64/include/asm/smp.h | 7 +------
>  1 file changed, 1 insertion(+), 6 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/smp.h b/arch/arm64/include/asm/smp.h
> index a62db952ffcb..d514383d6219 100644
> --- a/arch/arm64/include/asm/smp.h
> +++ b/arch/arm64/include/asm/smp.h
> @@ -37,12 +37,7 @@
>  
>  DECLARE_PER_CPU_READ_MOSTLY(int, cpu_number);
>  
> -/*
> - * We don't use this_cpu_read(cpu_number) as that has implicit writes to
> - * preempt_count, and associated (compiler) barriers, that we'd like to avoid
> - * the expense of. If we're preemptible, the value can be stale at use anyway.
> - */
> -#define raw_smp_processor_id() (*this_cpu_ptr(&cpu_number))
> +#define raw_smp_processor_id() (this_cpu_read(cpu_number))

I think the issue here is that, in the case of CONFIG_DEBUG_PREEMPT=y,
this_cpu_ptr ends up calling back into raw_smp_processor_id() via
my_cpu_offset, whereas this_cpu_read always uses __my_cpu_offset and avoids
the loop.

The right answer is probably to use raw_cpu_ptr instead, and update the
comment to explain why. Do you have CONFIG_DEBUG_PREEMPT=y?

Will